Dimensions of volume and segmentations are not matching

[stefano-trebeschi]

Following on the same public dataset of #512
I get non-matching dimensions between the CT volumes and the labels after running segimage2itkimage

[fedorov]

This is expected, and it is a very common question.

DICOM segmentations are represented as a set of frames (slices) - not as a single 3D volume. Those individual frames include geometry information that can be used to order them and reconstruct the segmented volume. For the sake of reducing file size while encoding SEG, frames that do not contain any segmented pixels are often skipped and are not included. As a result, when you have empty image slices above and/or below (which is often the case) the segmented slices, they will not be included.

When DICOM SEG is converted by segimage2itkimage, only the slices present in SEG will appear in the output representation. If you want the segmentation geometry to match that of the segmented image, you will need to augment the segmentation image with the empty slices above/below, which is easiest done by resampling segmentation to the volume of the segmented image.

Does it make sense?

[stefano-trebeschi]

Resampling to match the voxel space, or literally adding empty slices on top and bottom? As I see it those are two different operations

…

On Mon, Jan 13, 2025 at 22:29 Andrey Fedorov ***@***.***> wrote: This is expected, and it is a very common question. DICOM segmentations are represented as a set of frames (slices) - not as a single 3D volume. Those individual frames include geometry information that can be used to order them and reconstruct the segmented volume. For the sake of reducing file size while encoding SEG, frames that do not contain any segmented pixels are often skipped and are not included. As a result, when you have empty image slices above and/or below (which is often the case) the segmented slices, they will not be included. When DICOM SEG is converted by segimage2itkimage, only the slices present in SEG will appear in the output representation. If you want the segmentation geometry to match that of the segmented image, you will need to augment the segmentation image with the empty slices above/below, which is easiest done by resampling segmentation to the volume of the segmented image. Does it make sense? — Reply to this email directly, view it on GitHub <#513 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AWDP6QAQHJDDIOYE7VIMT332KQV37AVCNFSM6AAAAABVDPF2VWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOBYGI2TKOBZGY> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[fedorov]

If you do resampling to the raster of the segmented image with nearest neighbor interpolator (using ITK, as an example), the result should be identical to adding empty slices, but you will not need to worry about the code being error-prone while accounting for differences in image orientation (perhaps among other possible complexities).